Natural Language Generation

Natural Language Generation (NLG) is the area of Natural Language Processing that focusses on the construction of computer programs that produce text. This can be done for practical purposes (for example to perform data summarisation), or to simulate the human language production capability. 

Until recently, NLG was often regarded as the poor cousin of Natural Language Understanding. This started to change around 2010, when more and more researchers started to work on NLG. Since the Autumn of 2022, “generative AI” (e.g., prompt-diven models such as ChatGPT) is attracting huge attention, combining years of progress in NLG and other areas of NLP.

The Utrecht NLP has a strong tradition of work in, among others, the following areas of NLG.

1. Referring Expressions Generation. We construct and evaluate computer models that mimic human use of referring expressions. See Kees van Deemter’s book Computational Models of Referring for an overview, and see this paper by Fahime Same and others for one of our recent contributions.

2. Vision and Language Modelling.

3. Expressing logical information. A long tradition of work is engaged with constructing models that generate sentences from formulas of Mathematical Logic. As part of this tradition, Eduardo Calo’s PhD project (part of the NL4XAI project) is, given a logical formula, to verbalise the information contained in that formula (i.e., its truth conditions) with optimal clarity. A related branch of our work investigates how human speakers use quantifiers and other logical constructs, as in this recent paper in the journal JAIR.

4. Multilinguality. Although much of our NLG work focusses on English, we also work on other languages, such as Maltese (by Albert Gatt), and Mandarin (see e.g. Guanyi Chen’s thesis).

5. Methological issues. We’re interested in methodological issues arising from NLG and other areas of NLP. A recent example is this Squib in the Computational Linguistics journal, which proposes a framework for assessing the explanatory power of NLP models. Another is Yupei Du’s  ACL paper on instability in neural models.